tyandzx / caffe

Triplet loss for Caffe

Introduce triplet loss layer to caffe.
Concretely, we use cosine matric to constrain the distance between samples among same label/different labels.

Useage

1st, you need create a ordered file list for training.
This file list control the exactly data read-in order during training phase. Suppose in each mini-batch, you have data from 4 labels and 2 samples in each label, then the content of the file list should be like this:

  img_path_1_from_label_1 label_1
  img_path_2_from_label_1 label_1
  img_path_1_from_label_2 label_2
  img_path_2_from_label_2 label_2
  ...
  img_path_1_from_label_4 label_4
  img_path_2_from_label_4 label_4

2nd, define network structure in your train_val.prototxt.
Setup SampleTripletLayer to sample triplets in each mini-batch. Currently, triplets are made up by all anchor-positive pairs in the same label and one hardest negative sample from other labels.

  layer {
    name: "sample_triplet"
    type: "SampleTriplet"
    bottom: "fully_connected_feature"
    top: "triplet"
    sample_triplet_param {
      label_num: 4
      sample_num: 2
    }
  }

Setup TripletLossLayer to calculate loss.

  layer {
    name: "triplet_loss"
    type: "TripletLoss"
    bottom: "fully_connected_feature"
    bottom: "triplet"
    top: "triplet_loss"
    triplet_loss_param {
      margin: 0.1
    }
    loss_weight: 1
  }

caffe's Issues

关于测试集的batch大小设置

你好，非常感谢分享这个实现了triplet loss的代码。
我的问题是测试集的batch大小是否也要设置得跟训练集的batch大小一样？或者说，在测试的时候，测试集是否也需要实现triplet loss？

Changing condition

The condition should not been this(?):

dist_ap - dist_an + margin < 0

caffe/src/caffe/layers/triplet_loss_layer.cpp

Line 65 in 4cc936e

if (dist_ap - dist_an - margin < 0) {

about the triplet loss layer

in your example,the triplet loss layer has two bottom, but it need 3 bottoms as input, can u give a complete example. thanks very much.

FaceNet architecture

Hi, I am trying to implment the FaceNet architecture using your triplet loss. I got it to running but the loss is unusually high (e.g. loss = 7.24332e+07). Could you check the following prototxt and see if I am using your tripletLoss and SampleTriplet correctly for this case. Thanks.

`name: "FaceNet"
layer {
name: "facenet"
type: "Data"
top: "data"
top: "label"
include {
phase: TRAIN
}
transform_param {
scale: 0.00390625
crop_size: 220
}
data_param {
source: "/media/esb172/Hard_Disk_2/facenet_data/tightface/webfaces_train_lmdb/"
batch_size: 44
backend: LMDB
}
}
layer {
name: "facenet"
type: "Data"
top: "data"
top: "label"
include {
phase: TEST
}
transform_param {
scale: 0.00390625
crop_size: 220
}
data_param {
source: "/media/esb172/Hard_Disk_2/facenet_data/tightface/webfaces_val_lmdb"
batch_size: 44
backend: LMDB
}
}
layer {
name: "conv1"
type: "Convolution"
bottom: "data"
top: "conv1"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 64
kernel_size: 7
stride: 2
pad: 3
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "pool1"
type: "Pooling"
bottom: "conv1"
top: "pool1"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
}
}
layer {
name: "rnorm1"
type: "LRN"
bottom: "pool1"
top: "rnorm1"
lrn_param {
local_size: 5
alpha: 1.0
beta: 0.75
}
}
layer {
name: "conv2a"
type: "Convolution"
bottom: "rnorm1"
top: "conv2a"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 64
kernel_size: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "conv2"
type: "Convolution"
bottom: "conv2a"
top: "conv2"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 192
kernel_size: 3
stride: 1
pad: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "rnorm2"
type: "LRN"
bottom: "conv2"
top: "rnorm2"
lrn_param {
local_size: 5
alpha: 1.0
beta: 0.75
}
}
layer {
name: "pool2"
type: "Pooling"
bottom: "rnorm2"
top: "pool2"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
pad: 1
}
}

layer {
name: "conv3a"
type: "Convolution"
bottom: "pool2"
top: "conv3a"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 192
kernel_size: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "conv3"
type: "Convolution"
bottom: "conv3a"
top: "conv3"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 384
kernel_size: 3
stride: 1
pad: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "pool3"
type: "Pooling"
bottom: "conv3"
top: "pool3"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
#pad: 1
}
}

layer {
name: "conv4a"
type: "Convolution"
bottom: "pool3"
top: "conv4a"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 384
kernel_size: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "conv4"
type: "Convolution"
bottom: "conv4a"
top: "conv4"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 256
kernel_size: 3
stride: 1
pad: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "conv5a"
type: "Convolution"
bottom: "conv4"
top: "conv5a"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 256
kernel_size: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "conv5"
type: "Convolution"
bottom: "conv5a"
top: "conv5"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 256
kernel_size: 3
stride: 1
pad: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "conv6a"
type: "Convolution"
bottom: "conv5"
top: "conv6a"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 256
kernel_size: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "conv6"
type: "Convolution"
bottom: "conv6a"
top: "conv6"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
convolution_param {
num_output: 256
kernel_size: 3
stride: 1
pad: 1
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "pool4"
type: "Pooling"
bottom: "conv6"
top: "pool4"
pooling_param {
pool: MAX
kernel_size: 3
stride: 2
#pad: 1
}
}

layer {
name: "fc1a"
type: "InnerProduct"
bottom: "pool4"
top: "fc1a"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "fc1b"
type: "InnerProduct"
bottom: "pool4"
top: "fc1b"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "maxout1"
type: "Eltwise"
bottom: "fc1a"
bottom: "fc1b"
top: "maxout1"
eltwise_param {
operation:MAX
}
}

layer {
name: "fc2a"
type: "InnerProduct"
bottom: "maxout1"
top: "fc2a"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "fc2b"
type: "InnerProduct"
bottom: "maxout1"
top: "fc2b"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
inner_product_param {
num_output: 4096
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "maxout2"
type: "Eltwise"
bottom: "fc2a"
bottom: "fc2b"
top: "maxout2"
eltwise_param {
operation:MAX
}
}

layer {
name: "fc7128"
type: "InnerProduct"
bottom: "maxout2"
top: "fc7128"
param {
lr_mult: 1
}
param {
lr_mult: 2
}
inner_product_param {
num_output: 128
weight_filler {
type: "gaussian"
std: 0.1
}
bias_filler {
type: "constant"
}
}
}

layer {
name: "sample_triplet"
type: "SampleTriplet"
bottom: "fc7128"
top: "sample_triplet"
sample_triplet_param {
label_num: 10575
sample_num: 2
}
}

layer {
name: "triplet_loss"
type: "TripletLoss"
bottom: "fc7128"
bottom: "sample_triplet"
bottom: "label"
top: "triplet_loss"
triplet_loss_param {
margin: 0.1
}
loss_weight: 1
}

`

多GPU时候应该怎样设置

您好，我现在正在试验您的代码，是在拥有100W人的数据集上微调一个已经训练好的人脸识别模型，我目前遇到的困惑是：只能使用单张显卡，不知道在多GPU下怎样设置您的TripletLoss层，我尝试用两张显卡，但似乎loss只在第一张上起作用。

make all error

您好。我修改Makefile.config了，设置CPU_ONLY=1，以及修改了Python，opencv，openblas等路径。同样的Makefile.config，在使用make all编译BVLC-caffe的时候没有问题。但编译该版本caffe时出现如下错误：（还未找到解决方法，在这里求助一下）
.......
.......
CXX examples/cifar10/convert_cifar_data.cpp
CXX .build_debug/src/caffe/proto/caffe.pb.cc
AR -o .build_debug/lib/libcaffe.a
LD -o .build_debug/lib/libcaffe.so.1.0.0-rc3
CXX/LD -o .build_debug/tools/test_net.bin
CXX/LD -o .build_debug/tools/train_net.bin
CXX/LD -o .build_debug/tools/upgrade_net_proto_binary.bin
CXX/LD -o .build_debug/tools/upgrade_solver_proto_text.bin
CXX/LD -o .build_debug/tools/device_query.bin
CXX/LD -o .build_debug/tools/finetune_net.bin
CXX/LD -o .build_debug/tools/convert_imageset.bin
CXX/LD -o .build_debug/tools/extract_features.bin
..build_debug/build_debug/lib/liblibcaffe.so�/libcaffe.so����寚�‘caffe::SampleTripletLayer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe��::Blob>� >� const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob��caffe::,SampleTripletLayer:std::allocator<:caffe::BlobBackward_gpu<double(>>std > const&):��:�未定义的引用vector<caffe::Blob<double

,. build_debug/stdlib/:libcaffe.so：对:‘allocatorcaffe::TripletLossLayer<:caffe:Forward_gpu(std:::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<:caffe::Blob, stdBlob::allocator<caffe:<:Blobdouble>> > const&)’未定义的引用
.build_debug/>lib/libcaffe.so：对� ��caffe::>SampleTripletLayer< double>::Forward_gpu(stdconst::&vector<,caffe: :Blob, std::stdallocator<:caffe::vector<bool, std:::allocator< > doubleconst&>, std::vector<caffe::Blob**, >std: :allocatorcaffe::Blob< double>const> &> , std::constvector<caffe&::)�Blob<double�>�,� std�::�allocator<�caffe::Blob<��double>> > �const&)‹�未定义的引甉�
�.build_debug�/lib�/libcaffe.so��：对╀�caffe::TripletLossLayer::Backward_gpu��(std::vector<caffe::Blob�<float
,. stdbuild_debug/::allocator<caffe::Blob> > constlib/&libcaffe.so：��,�� ��caffestd:::TripletLossLayer<:float>vector<bool::,Forward_gpu(std::vector<caffe:: stdBlob, std::allocator<::caffe::Blob> allocator> const&, std:<:vectorboolcaffe::Blob, std::allocator<caffe::Blob> > const&)’ >未定� const��的引用
.build_debug/lib/libcaffe.so：对‘caffe&::,SampleTripletLayer::Forward_gpu(std::vector<caffe:: Blob, std::allocator<caffe::Blobstd:> > const&, std:::vectorvector<caffe::Blob, std::allocator<<caffe:caffe:Blob:double:> > const&)’未定义的引用
.build_debug/lib/Blob<libcaffe.so：对‘caffe::TripletLossLayerfloat>::Backward_gpu(std::vector<caffe::Blob**, ,std: :allocatorstd<:caffe::Blobfloat:allocator<> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<::caffe::Blob> > const&)’未定�Blob��的引用
.<floatbuild_debug/lib/libcaffe.so：对‘caffe::SampleTripletLayer:>:Forward_gpu(std>:: vector<caffe::Blob,> std ::allocatorconst<caffe&::)Blob�> �>� �const&�, �std::vector<caffe::Blob��, std::allocator<caffe::Blob> > const&)’未定乚�的引用
.build_debug��/lib/libcaffe.so：对‘caffe::TripletLossLayer:�:Forward_gpu(std:�:vector��<caffe::Blob�, �std�:�:�allocator<caffe::�Blob
.build_debug > const&, std:/:vector<caffe:lib/:Blob<libcaffe.sodouble>, �std::allocator��<caffe::Blob�>� �>� �const&)’未定义的引用
.build_debug/lib/libcaffe.so�caffe：对‘caffe::SampleTripletLayerfloat::::Backward_gpu(std:SampleTripletLayer:vector<caffe:<:float>Blob<::float>, std::allocator<caffe:Forward_gpu:Blob<float(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::TripletLossLayer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob>> > const&)’未定义的引用
.build_debug/lib> /libcaffe.so：对‘caffe::SampleTripletLayer>:: Backward_gpu(stdconst::&vector,< caffe:std:Blob, std::allocator<caffe::::Blob<vector> > const&, std::vector<bool, boolstd::allocator, > const&, stdstd:::vector<caffe:::Blob<, std:bool>:allocator<caffe::Blob>> > const&const&)’未定义的引, 用
.build_debug/lib/libcaffe.so：std��::�‘caffe::TripletLossLayer:<:Backward_gpucaffe:(std::vector<caffe::Blob, std::allocator<caffe:::Blob<Blob> > const&, std::floatvector<bool, std:>:allocator, std::<allocator > :const&,:Blob std> >: const&:vector)’未定义的引用
<.build_debug/caffe:lib/:Bloblibcaffe.so：对‘<caffe::TripletLossLayer:>, std:::allocator<caffeBackward_gpu(::stdBlob<:double>>: > constvector<&)��caffe:�未定义的引用
:Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocatorcollect2: error: ld returned 1 exit status
> const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
collect2: error: ld returned 1 exit status
make: *** [.build_debug/tools/upgrade_solver_proto_text.bin] 错误 1
make: *** 正在等待未完成的任务....
make: *** [.build_debug/tools/upgrade_net_proto_binary.bin] 错误 1
.build_debug/lib/libcaffe.so：对‘caffe::SampleTripletLayer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::TripletLossLayer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::SampleTripletLayer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::TripletLossLayer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::SampleTripletLayer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::TripletLossLayer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::SampleTripletLayer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::TripletLossLayer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
collect2: error: ld returned 1 exit status
make: *** [.build_debug/tools/convert_imageset.bin] 错误 1
.build_debug/lib/libcaffe.so：对‘caffe::SampleTripletLayer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::TripletLossLayer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::SampleTripletLayer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::TripletLossLayer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::SampleTripletLayer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::TripletLossLayer::Forward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::SampleTripletLayer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
.build_debug/lib/libcaffe.so：对‘caffe::TripletLossLayer::Backward_gpu(std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&, std::vector<bool, std::allocator > const&, std::vector<caffe::Blob, std::allocator<caffe::Blob> > const&)’未定义的引用
collect2: error: ld returned 1 exit status
make: *** [.build_debug/tools/extract_features.bin] 错误 1

major reference ?

Hi,
I want to know your reference(paper) about triplet loss.
Can you give me the paper you refer to? facenet??

loss 在0.7500上下跳动

您好，我由于项目需要，临时了解了一些尺度学习的知识，所以感觉用起来有很多地方并不了解，所以想请教一下。

1. 我看前面的issue有说，您采用的是三角的loss，不是很懂这个，我还以为是l2的loss，请问下这个loss的全称或者公式？

2.我们一开始觉得可能和别的任务一样，深层网络较好，所以用了Res100和Res50去训练，但是loss一直停留在0.75（margin 0.25），不是波动，而是保持在

3. 我们最后采用了ZF-net结构
   `name:"ZF car detect"
   layer {
   name: "data"
   type: "ImageData"
   top: "data"
   top: "label"
   transform_param {
   mean_value: 127.5
   mean_value: 127.5
   mean_value: 127.5
   scale: 0.0078125

mirror: true

}
image_data_param {
new_height: 256
new_width: 256
source: "imagesets/train.txt"
batch_size: 32
shuffle: false
}
}

layer {
name: "conv1"
type: "Convolution"
bottom: "data"
top: "conv1"
param { lr_mult: 1.0 }
param { lr_mult: 2.0 }
convolution_param {
num_output: 96
kernel_size: 7
stride: 2
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0
}
}
}

layer {
name: "relu1"
type: "ReLU"
bottom: "conv1"
top: "conv1"
}

layer {
name: "pool1"
type: "Pooling"
bottom: "conv1"
top: "pool1"
pooling_param {
kernel_size: 3
stride: 2
pool: MAX
}
}

layer {
name: "conv2"
type: "Convolution"
bottom: "pool1"
top: "conv2"
param { lr_mult: 1.0 }
param { lr_mult: 2.0 }
convolution_param {
num_output: 256
kernel_size: 5
stride: 2
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0
}
}
}

layer {
name: "relu2"
type: "ReLU"
bottom: "conv2"
top: "conv2"
}

layer {
name: "pool2"
type: "Pooling"
bottom: "conv2"
top: "pool2"
pooling_param {
kernel_size: 3
stride: 2
pool: MAX
}
}

layer {
name: "conv3"
type: "Convolution"
bottom: "pool2"
top: "conv3"
param { lr_mult: 1.0 }
param { lr_mult: 2.0 }
convolution_param {
num_output: 512
kernel_size: 3
stride: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0
}
}
}
...
some other layers
...
layer {
name:"fc8"
bottom: "fc7"
top: "fc8"
type: "InnerProduct"
inner_product_param {
num_output: 512
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
value: 0
}
}
}

layer {
name: "sample_triplet"
type: "SampleTriplet"
bottom: "fc8"
top: "triplet"
sample_triplet_param {
label_num: 8
sample_num: 4
}
}

layer {
name: "triplet_loss"
type: "TripletLoss"
bottom: "fc8"
bottom: "triplet"
top: "triplet_loss"
triplet_loss_param {
margin: 0.25
}
loss_weight: 1
}
训练一晚上后的结果，从开始到最后一直在0.75的小数点后4位波动
I0614 17:36:31.718358 28781 solver.cpp:244] Train net output #32: triplet_loss = 0.749998 (* 1 = 0.749998 loss)
I0614 17:36:31.718369 28781 sgd_solver.cpp:106] Iteration 446000, lr = 1e-08
I0614 17:36:34.979749 28781 solver.cpp:228] Iteration 446020, loss = 0.75007
I0614 17:36:35.046409 28781 solver.cpp:244] Train net output #0: label = 14537
I0614 17:36:35.046422 28781 solver.cpp:244] Train net output #1: label = 14537
I0614 17:36:35.046438 28781 solver.cpp:244] Train net output #2: label = 14537
I0614 17:36:35.046443 28781 solver.cpp:244] Train net output #3: label = 14537
I0614 17:36:35.046448 28781 solver.cpp:244] Train net output #4: label = 12832
I0614 17:36:35.046453 28781 solver.cpp:244] Train net output #5: label = 12832
I0614 17:36:35.046458 28781 solver.cpp:244] Train net output #6: label = 12832
I0614 17:36:35.046468 28781 solver.cpp:244] Train net output #7: label = 12832
I0614 17:36:35.046478 28781 solver.cpp:244] Train net output #8: label = 12280
I0614 17:36:35.046488 28781 solver.cpp:244] Train net output #9: label = 12280
I0614 17:36:35.046499 28781 solver.cpp:244] Train net output #10: label = 12280
I0614 17:36:35.046509 28781 solver.cpp:244] Train net output #11: label = 12280
I0614 17:36:35.046519 28781 solver.cpp:244] Train net output #12: label = 27357
I0614 17:36:35.046530 28781 solver.cpp:244] Train net output #13: label = 27357
I0614 17:36:35.046540 28781 solver.cpp:244] Train net output #14: label = 27357
I0614 17:36:35.046550 28781 solver.cpp:244] Train net output #15: label = 27357
I0614 17:36:35.046561 28781 solver.cpp:244] Train net output #16: label = 7992
I0614 17:36:35.046571 28781 solver.cpp:244] Train net output #17: label = 7992
I0614 17:36:35.046581 28781 solver.cpp:244] Train net output #18: label = 7992
I0614 17:36:35.046591 28781 solver.cpp:244] Train net output #19: label = 7992
I0614 17:36:35.046602 28781 solver.cpp:244] Train net output #20: label = 9613
I0614 17:36:35.046613 28781 solver.cpp:244] Train net output #21: label = 9613
I0614 17:36:35.046623 28781 solver.cpp:244] Train net output #22: label = 9613
I0614 17:36:35.046634 28781 solver.cpp:244] Train net output #23: label = 9613
I0614 17:36:35.046645 28781 solver.cpp:244] Train net output #24: label = 19435
I0614 17:36:35.046655 28781 solver.cpp:244] Train net output #25: label = 19435
I0614 17:36:35.046665 28781 solver.cpp:244] Train net output #26: label = 19435
I0614 17:36:35.046672 28781 solver.cpp:244] Train net output #27: label = 19435
I0614 17:36:35.046682 28781 solver.cpp:244] Train net output #28: label = 219
I0614 17:36:35.046694 28781 solver.cpp:244] Train net output #29: label = 219
I0614 17:36:35.046703 28781 solver.cpp:244] Train net output #30: label = 219
I0614 17:36:35.046715 28781 solver.cpp:244] Train net output #31: label = 219
I0614 17:36:35.046730 28781 solver.cpp:244] Train net output #32: triplet_loss = 0.75007 (* 1 = 0.75007 loss)
I0614 17:36:35.046741 28781 sgd_solver.cpp:106] Iteration 446020, lr = 1e-08
I0614 17:36:38.170933 28781 solver.cpp:228] Iteration 446040, loss = 0.750116
I0614 17:36:38.238198 28781 solver.cpp:244] Train net output #0: label = 384
I0614 17:36:38.238209 28781 solver.cpp:244] Train net output #1: label = 384
I0614 17:36:38.238226 28781 solver.cpp:244] Train net output #2: label = 384
I0614 17:36:38.238231 28781 solver.cpp:244] Train net output #3: label = 384
I0614 17:36:38.238235 28781 solver.cpp:244] Train net output #4: label = 11448
I0614 17:36:38.238240 28781 solver.cpp:244] Train net output #5: label = 11448
I0614 17:36:38.238246 28781 solver.cpp:244] Train net output #6: label = 11448
I0614 17:36:38.238256 28781 solver.cpp:244] Train net output #7: label = 11448
I0614 17:36:38.238266 28781 solver.cpp:244] Train net output #8: label = 9058
I0614 17:36:38.238276 28781 solver.cpp:244] Train net output #9: label = 9058
I0614 17:36:38.238287 28781 solver.cpp:244] Train net output #10: label = 9058
I0614 17:36:38.238297 28781 solver.cpp:244] Train net output #11: label = 9058
I0614 17:36:38.238308 28781 solver.cpp:244] Train net output #12: label = 26239
I0614 17:36:38.238319 28781 solver.cpp:244] Train net output #13: label = 26239
I0614 17:36:38.238348 28781 solver.cpp:244] Train net output #14: label = 26239
I0614 17:36:38.238359 28781 solver.cpp:244] Train net output #15: label = 26239
I0614 17:36:38.238369 28781 solver.cpp:244] Train net output #16: label = 15103
I0614 17:36:38.238380 28781 solver.cpp:244] Train net output #17: label = 15103
I0614 17:36:38.238390 28781 solver.cpp:244] Train net output #18: label = 15103
I0614 17:36:38.238401 28781 solver.cpp:244] Train net output #19: label = 15103
I0614 17:36:38.238411 28781 solver.cpp:244] Train net output #20: label = 4531
I0614 17:36:38.238421 28781 solver.cpp:244] Train net output #21: label = 4531
I0614 17:36:38.238432 28781 solver.cpp:244] Train net output #22: label = 4531
I0614 17:36:38.238442 28781 solver.cpp:244] Train net output #23: label = 4531
I0614 17:36:38.238452 28781 solver.cpp:244] Train net output #24: label = 8527
I0614 17:36:38.238463 28781 solver.cpp:244] Train net output #25: label = 8527
I0614 17:36:38.238473 28781 solver.cpp:244] Train net output #26: label = 8527
I0614 17:36:38.238484 28781 solver.cpp:244] Train net output #27: label = 8527
I0614 17:36:38.238494 28781 solver.cpp:244] Train net output #28: label = 17737
I0614 17:36:38.238505 28781 solver.cpp:244] Train net output #29: label = 17737
I0614 17:36:38.238515 28781 solver.cpp:244] Train net output #30: label = 17737
I0614 17:36:38.238525 28781 solver.cpp:244] Train net output #31: label = 17737
I0614 17:36:38.238540 28781 solver.cpp:244] Train net output #32: triplet_loss = 0.750116 (* 1 = 0.750116 loss)
I0614 17:36:38.238554 28781 sgd_solver.cpp:106] Iteration 446040, lr = 1e-08

`

训练数据的话有约3k个类。不过我们的图像比较凌乱，大小不一，没什么预处理，不知道这会不会有影响。
谢谢！

[error] Check failed: ExactNumBottomBlobs() == bottom.size() (2 vs. 3) TripletLoss Layer takes 2 bottom blob(s) as input.

hi.
i try your triplet loss. but i have next problem.

F0721 12:28:31.108945 19808 layer.hpp:374] Check failed: ExactNumBottomBlobs() == bottom.size() (2 vs. 3) TripletLoss Layer takes 2 bottom blob(s) as input.
*** Check failure stack trace: ***
@ 0x7fa8885d0b5d google::LogMessage::Fail()
@ 0x7fa8885d4b77 google::LogMessage::SendToLog()
@ 0x7fa8885d29f9 google::LogMessage::Flush()
@ 0x7fa8885d2cfd google::LogMessageFatal::~LogMessageFatal()
@ 0x7fa88e3e1f01 caffe::Layer<>::CheckBlobCounts()
@ 0x7fa88e42de13 caffe::Net<>::Init()
@ 0x7fa88e43028b caffe::Net<>::Net()
@ 0x7fa88e50778a caffe::Solver<>::InitTrainNet()
@ 0x7fa88e507e74 caffe::Solver<>::Init()
@ 0x7fa88e5083a8 caffe::Solver<>::Solver()
@ 0x7fa88e409d13 caffe::Creator_SGDSolver<>()
@ 0x412193 caffe::SolverRegistry<>::CreateSolver()
@ 0x40c0c8 train()
@ 0x4089af main
@ 0x33f3a1ed5d (unknown)
@ 0x407ef9 (unknown)
중지됨

Is triplet loss need 3 input size ?

Sample data-label file and model prototxt

Hi,

I'm trying to use your triplet loss for cifar10. However my train loss isn't decreasing, and my validation accuracy is always 0. I believe the problem is with my input.

Can you share an example model along with image-label (for certain batch size) file which will help us understand what is required.

Thank you

训练准确率一直不变

我按照说明种的图片排列和网络配置方式，训练了50个类，每个类2张图片，准确率一直不变，为随机初始的值，因为readme中没有详细的教程，所以不知是否配置错误还是其他

about the negative node selection in triplets samples

In your triplets sampling layer, you will firstly select the <anchor, positive>, when to find the best negative to form the triplets, but I see in your code(https://github.com/tyandzx/caffe/blob/master/src/caffe/layers/sample_triplet_layer.cpp#L48), you will select the negative node that is most far from the anchor, not the closest one (hardest negative) or using semi-hardest negative proposed in FaceNet, like

What is your considerations? thanks.

About label_num and sample_num

Hello @tyandzx.
In your example, we have 4 label_num and 2 sample_num each label_num.
The problem is in my data, number of each label is allway different.
Whether:

1. I can igorne two these parameter.
2. I can make a lits train with number of each is different.
   example:

in train.prototxt

layer {
name: "sample_triplet"
type: "SampleTriplet"
bottom: "loss3/classifier"
top: "triplet"
}
}

in listtrain

data\facenet\images\Aaron_Peirsol_0002.jpg 1
data\facenet\images\Aaron_Peirsol_0003.jpg 1
data\facenet\images\Aaron_Peirsol_0004.jpg 1
data\facenet\images\Aaron_Sorkin_0002.jpg 2
data\facenet\images\Abdoulaye_Wade_0002.jpg 3
data\facenet\images\Abdoulaye_Wade_0003.jpg 3
data\facenet\images\Abdoulaye_Wade_0004.jpg 3
data\facenet\images\Abdullah_0002.jpg 4
data\facenet\images\Abdullah_0003.jpg 4
data\facenet\images\Abdullah_0004.jpg 4
data\facenet\images\Abdullah_al-Attiyah_0002.jpg 5
data\facenet\images\Abdullah_al-Attiyah_0003.jpg 5

Thank you so much